search

xingyizhou / pytorch-pose-hg-3d

PyTorch implementation for 3D human pose estimation
GNU General Public License v3.0
614 stars 143 forks source link

Increasing accuracy #59

Open Kyanite123 opened 5 years ago

Kyanite123 commented 5 years ago

Hi,

Thanks for sharing with the community your code! This is very helpful!

I tried to improve the accuracy of the estimated results by increasing the following parameters to: outputRes = 128 inputRes = 512

But now it seems that the network cannot converge to a reasonable result at all.

Can you please share some insights of why this might happen and possible what parameters to adjust?

Thanks!

xingyizhou commented 5 years ago

Hi, Thanks very much for the extension of the project. I haven't tried using higher resolution so I don't have a 100% sure solution. According to my experiences, when changing the resolution, you should also change opt.hm_res so that the ratio of positive and negative samples doesn't change (following Xiao et el https://github.com/Microsoft/human-pose-estimation.pytorch/blob/master/experiments/mpii/resnet50/384x384_d256x3_adam_lr1e-3.yaml#L28 https://github.com/Microsoft/human-pose-estimation.pytorch/blob/master/experiments/mpii/resnet50/256x256_d256x3_adam_lr1e-3.yaml#L28). However, since the original resolution of single human pose estimation is not as large as 512x512, I am not sure if using higher resolution will help a lot.